GBA-LIOM: Global Bundle Adjustment Enhanced LiDAR–Inertial Odometry and Mapping

An accurate and consistent LiDAR point cloud map has a wide range of applications, such as in building information modeling, online retailing, and autonomous vehicles. However, existing methods face various drawbacks and challenges. First, odometry divergence often occurs in degraded, narrow corridors due to the lack of features. In addition, large-scale global optimization with LiDAR point clouds remains challenging because of memory limitations and the complex derivations involved in the process. Toward removing these barriers, this article proposes a consistent and efficient large-scale LiDAR odometry and mapping method. First, we use distribution-to-multidistribution constraints in odometry to deal with challenging scenes. Second, we propose a novel storage structure for LiDAR features, significantly reducing the excessive memory consumption of existing optimization methods. Moreover, we reformulate the error term in the optimization problem, enabling the use of generic least squares solvers and eliminating the need for complex derivative derivations. These improvements result in a substantial acceleration of LiDAR bundle adjustment (BA). Third, the optimization focuses on minimizing the roughness of surfels within windows of selected adjacent LiDAR frames. By assigning frames around the loop to a window, our method naturally integrates loop closure into the optimization. This allows our method to effectively incorporate loop constraints in BA. Finally, the proposed method is validated in terms of memory efficiency, computational speed, mapping consistency, and accuracy using a public benchmark and our own datasets.